CPUPowerWatcher/Seer: Automated and Flexible CPU Power Modelling

Abstract

Power supply is a key limitation when scaling supercomputing capabilities, making power consumption a major challenge in HPC field. To develop energy-efficient tools, it is essential to have an accurate power consumption modelling. Although previous works proposed several approaches to model CPU power, building models in an automated and adaptable way, and accurately predicting power, remains complex. This work presents two tools: CPUPowerWatcher, which gathers CPU metrics during the execution of user-defined workloads, and CPUPowerSeer, which build models to predict power from time series data. Using both tools, experiments were conducted to analyse the impact of novel factors on CPU power and compare the accuracy of six regression models when predicting CPU- and I/O-intensive workloads.